mosesdecoder/moses/LM/DALMWrapper.cpp

//#include <boost/functional/hash.hpp>
#include <algorithm>
#include "moses/FF/FFState.h"
#include "DALMWrapper.h"
#include "dalm.h"
#include "moses/FactorTypeSet.h"
#include "moses/FactorCollection.h"
#include "moses/InputFileStream.h"
#include "util/exception.hh"
#include "moses/ChartHypothesis.h"
#include "moses/ChartManager.h"

using namespace std;

/////////////////////////
void read_ini(const char *inifile, string &model, string &words, string &wordstxt)
{
  ifstream ifs(inifile);
  string line;

  getline(ifs, line);
  while(ifs) {
    unsigned int pos = line.find("=");
    string key = line.substr(0, pos);
    string value = line.substr(pos+1, line.size()-pos);
    if(key=="MODEL") {
      model = value;
    } else if(key=="WORDS") {
      words = value;
    } else if(key=="WORDSTXT") {
      wordstxt = value;
    }
    getline(ifs, line);
  }
}
/////////////////////////

namespace Moses
{

class Murmur: public DALM::State::HashFunction
{
public:
  Murmur(std::size_t seed=0): seed(seed) {
  }
  virtual std::size_t operator()(const DALM::VocabId *words, std::size_t size) const {
    return util::MurmurHashNative(words, sizeof(DALM::VocabId) * size, seed);
  }
private:
  std::size_t seed;
};

class DALMState : public FFState
{
private:
  DALM::State state;

public:
  DALMState() {
  }

  DALMState(const DALMState &from) {
    state = from.state;
  }

  virtual ~DALMState() {
  }

  void reset(const DALMState &from) {
    state = from.state;
  }

  virtual int Compare(const FFState& other) const {
    const DALMState &o = static_cast<const DALMState &>(other);
    if(state.get_count() < o.state.get_count()) return -1;
    else if(state.get_count() > o.state.get_count()) return 1;
    else return state.compare(o.state);
  }

  virtual size_t hash() const {
    // imitate KenLM
    return state.hash(Murmur());
  }

  virtual bool operator==(const FFState& other) const {
    const DALMState &o = static_cast<const DALMState &>(other);
    return state.compare(o.state) == 0;
  }

  DALM::State &get_state() {
    return state;
  }

  void refresh() {
    state.refresh();
  }
};

class DALMChartState : public FFState
{
private:
  DALM::Fragment prefixFragments[DALM_MAX_ORDER-1];
  unsigned char prefixLength;
  DALM::State rightContext;
  bool isLarge;

public:
  DALMChartState()
    : prefixLength(0),
      isLarge(false) {
  }

  virtual ~DALMChartState() {
  }

  inline unsigned char GetPrefixLength() const {
    return prefixLength;
  }

  inline unsigned char &GetPrefixLength() {
    return prefixLength;
  }

  inline const DALM::Fragment *GetPrefixFragments() const {
    return prefixFragments;
  }

  inline DALM::Fragment *GetPrefixFragments() {
    return prefixFragments;
  }

  inline const DALM::State &GetRightContext() const {
    return rightContext;
  }

  inline DALM::State &GetRightContext() {
    return rightContext;
  }

  inline bool LargeEnough() const {
    return isLarge;
  }

  inline void SetAsLarge() {
    isLarge=true;
  }

  virtual int Compare(const FFState& other) const {
    const DALMChartState &o = static_cast<const DALMChartState &>(other);
    if(prefixLength < o.prefixLength) return -1;
    if(prefixLength > o.prefixLength) return 1;
    if(prefixLength!=0) {
      const DALM::Fragment &f = prefixFragments[prefixLength-1];
      const DALM::Fragment &of = o.prefixFragments[prefixLength-1];
      int ret = DALM::compare_fragments(f,of);
      if(ret != 0) return ret;
    }
    if(isLarge != o.isLarge) return (int)isLarge - (int)o.isLarge;
    if(rightContext.get_count() < o.rightContext.get_count()) return -1;
    if(rightContext.get_count() > o.rightContext.get_count()) return 1;
    return rightContext.compare(o.rightContext);
  }

  virtual size_t hash() const {
    // imitate KenLM
    unsigned char add[2];
    add[0] = prefixLength;
    add[1] = isLarge;
    std::size_t seed = util::MurmurHashNative(add, 2, prefixLength ? prefixFragments[prefixLength-1].sid : 0);
    return rightContext.hash(Murmur(seed));
  }

  virtual bool operator==(const FFState& other) const {
    const DALMChartState &o = static_cast<const DALMChartState &>(other);

    // check left state.
    if(prefixLength != o.prefixLength) return false;
    const DALM::Fragment &f = prefixFragments[prefixLength-1];
    const DALM::Fragment &of = o.prefixFragments[prefixLength-1];
    if(DALM::compare_fragments(f, of) != 0) return false;

    // check right state.
    if(rightContext.get_count() != o.rightContext.get_count()) return false;
    return rightContext.compare(o.rightContext) == 0;
  }

};

LanguageModelDALM::LanguageModelDALM(const std::string &line)
  :LanguageModel(line)
{
  ReadParameters();

  if (m_factorType == NOT_FOUND) {
    m_factorType = 0;
  }
}

LanguageModelDALM::~LanguageModelDALM()
{
  delete m_logger;
  delete m_vocab;
  delete m_lm;
}

void LanguageModelDALM::Load(AllOptions::ptr const& opts)
{
  /////////////////////
  // READING INIFILE //
  /////////////////////
  string inifile= m_filePath + "/dalm.ini";

  string model; // Path to the double-array file.
  string words; // Path to the vocabulary file.
  string wordstxt; //Path to the vocabulary file in text format.
  read_ini(inifile.c_str(), model, words, wordstxt);

  model = m_filePath + "/" + model;
  words = m_filePath + "/" + words;
  wordstxt = m_filePath + "/" + wordstxt;

  UTIL_THROW_IF(model.empty() || words.empty() || wordstxt.empty(),
                util::FileOpenException,
                "Failed to read DALM ini file " << m_filePath << ". Probably doesn't exist");

  ////////////////
  // LOADING LM //
  ////////////////

  // Preparing a logger object.
  m_logger = new DALM::Logger(stderr);
  m_logger->setLevel(DALM::LOGGER_INFO);

  // Load the vocabulary file.
  m_vocab = new DALM::Vocabulary(words, *m_logger);

  // Load the language model.
  m_lm = new DALM::LM(model, *m_vocab, m_nGramOrder, *m_logger);

  wid_start = m_vocab->lookup(BOS_);
  wid_end = m_vocab->lookup(EOS_);

  // vocab mapping
  CreateVocabMapping(wordstxt);

  FactorCollection &collection = FactorCollection::Instance();
  m_beginSentenceFactor = collection.AddFactor(BOS_);
}

const FFState *LanguageModelDALM::EmptyHypothesisState(const InputType &/*input*/) const
{
  DALMState *s = new DALMState();
  m_lm->init_state(s->get_state());
  return s;
}

void LanguageModelDALM::CalcScore(const Phrase &phrase, float &fullScore, float &ngramScore, size_t &oovCount) const
{
  oovCount = 0;
  fullScore  = 0.0f;
  ngramScore = 0.0f;

  size_t phraseSize = phrase.GetSize();
  if (!phraseSize) return;

  size_t currPos = 0;
  //size_t hist_count = 0;
  DALM::State state;

  if(phrase.GetWord(0).GetFactor(m_factorType) == m_beginSentenceFactor) {
    m_lm->init_state(state);
    currPos++;
    //hist_count++;
  }

  float score;
  float prefixScore=0.0f;
  float partScore=0.0f;
  //std::cerr << std::setprecision(8);
  //std::cerr << "# ";
  while (currPos < phraseSize) {
    const Word &word = phrase.GetWord(currPos);
    //hist_count++;

    if (word.IsNonTerminal()) {
      //std::cerr << "X ";
      state.refresh();
      fullScore += partScore;
      partScore = 0.0f;
      //std::cerr << fullScore << " ";
      //hist_count = 0;
    } else {
      //std::cerr << word.GetString(m_factorType).as_string() << " ";
      DALM::VocabId wid = GetVocabId(word.GetFactor(m_factorType));
      score = m_lm->query(wid, state);
      partScore += score;
      //std::cerr << partScore << " ";
      //if (hist_count >= m_nGramOrder) ngramScore += score;
      if (wid==m_vocab->unk()) ++oovCount;
    }

    currPos++;
    if (currPos >= m_ContextSize) {
      break;
    }
  }
  prefixScore = fullScore + partScore;
  //std::cerr << prefixScore << " ";

  while (currPos < phraseSize) {
    const Word &word = phrase.GetWord(currPos);
    //hist_count++;

    if (word.IsNonTerminal()) {
      //std::cerr << "X ";
      fullScore += partScore;
      partScore = 0.0f;
      //std::cerr << fullScore << " ";
      state.refresh();
      //hist_count = 0;
    } else {
      //std::cerr << word.GetString(m_factorType).as_string() << " ";
      DALM::VocabId wid = GetVocabId(word.GetFactor(m_factorType));
      score = m_lm->query(wid, state);
      partScore += score;
      //std::cerr << partScore << " ";
      if (wid==m_vocab->unk()) ++oovCount;
    }

    currPos++;
  }
  fullScore += partScore;

  ngramScore = TransformLMScore(fullScore - prefixScore);
  fullScore = TransformLMScore(fullScore);
}

FFState *LanguageModelDALM::EvaluateWhenApplied(const Hypothesis &hypo, const FFState *ps, ScoreComponentCollection *out) const
{
  // In this function, we only compute the LM scores of n-grams that overlap a
  // phrase boundary. Phrase-internal scores are taken directly from the
  // translation option.

  const DALMState *dalm_ps = static_cast<const DALMState *>(ps);

  // Empty phrase added? nothing to be done
  if (hypo.GetCurrTargetLength() == 0) {
    return dalm_ps ? new DALMState(*dalm_ps) : NULL;
  }

  const std::size_t begin = hypo.GetCurrTargetWordsRange().GetStartPos();
  //[begin, end) in STL-like fashion.
  const std::size_t end = hypo.GetCurrTargetWordsRange().GetEndPos() + 1;
  const std::size_t adjust_end = std::min(end, begin + m_nGramOrder - 1);

  DALMState *dalm_state = new DALMState(*dalm_ps);
  DALM::State &state = dalm_state->get_state();
  float score = 0.0;
  for(std::size_t position=begin; position < adjust_end; position++) {
    score += m_lm->query(GetVocabId(hypo.GetWord(position).GetFactor(m_factorType)), state);
  }

  if (hypo.IsSourceCompleted()) {
    // Score end of sentence.
    std::vector<DALM::VocabId> indices(m_nGramOrder-1);
    const DALM::VocabId *last = LastIDs(hypo, &indices.front());
    m_lm->set_state(&indices.front(), (last-&indices.front()), state);

    score += m_lm->query(wid_end, state);
  } else if (adjust_end < end) {
    // Get state after adding a long phrase.
    std::vector<DALM::VocabId> indices(m_nGramOrder-1);
    const DALM::VocabId *last = LastIDs(hypo, &indices.front());
    m_lm->set_state(&indices.front(), (last-&indices.front()), state);
  }

  score = TransformLMScore(score);
  if (OOVFeatureEnabled()) {
    std::vector<float> scores(2);
    scores[0] = score;
    scores[1] = 0.0;
    out->PlusEquals(this, scores);
  } else {
    out->PlusEquals(this, score);
  }

  return dalm_state;
}

FFState *LanguageModelDALM::EvaluateWhenApplied(const ChartHypothesis& hypo, int featureID, ScoreComponentCollection *out) const
{
  // initialize language model context state
  DALMChartState *newState = new DALMChartState();
  DALM::State &state = newState->GetRightContext();

  DALM::Fragment *prefixFragments = newState->GetPrefixFragments();
  unsigned char &prefixLength = newState->GetPrefixLength();

  // initial language model scores
  float hypoScore = 0.0;      // total hypothesis score.

  const TargetPhrase &targetPhrase = hypo.GetCurrTargetPhrase();
  size_t hypoSize = targetPhrase.GetSize();

  // get index map for underlying hypotheses
  const AlignmentInfo::NonTermIndexMap &nonTermIndexMap =
    targetPhrase.GetAlignNonTerm().GetNonTermIndexMap();

  size_t phrasePos = 0;

  // begginig of sentence.
  if(hypoSize > 0) {
    const Word &word = targetPhrase.GetWord(0);
    if(word.GetFactor(m_factorType) == m_beginSentenceFactor) {
      //std::cerr << word.GetString(m_factorType).as_string() << " ";
      m_lm->init_state(state);
      // state is finalized.
      newState->SetAsLarge();
      phrasePos++;
    } else if(word.IsNonTerminal()) {
      // special case: rule starts with non-terminal -> copy everything

      const ChartHypothesis *prevHypo = hypo.GetPrevHypo(nonTermIndexMap[0]);
      const DALMChartState* prevState =
        static_cast<const DALMChartState*>(prevHypo->GetFFState(featureID));


      // copy chart state
      (*newState) = (*prevState);

      phrasePos++;
    }
  }

  // loop over rule
  for (; phrasePos < hypoSize; phrasePos++) {

    // consult rule for either word or non-terminal
    const Word &word = targetPhrase.GetWord(phrasePos);

    // regular word
    if (!word.IsNonTerminal()) {
      EvaluateTerminal(
        word, hypoScore,
        newState, state,
        prefixFragments, prefixLength
      );
    }

    // non-terminal, add phrase from underlying hypothesis
    // internal non-terminal
    else {
      // look up underlying hypothesis
      const ChartHypothesis *prevHypo = hypo.GetPrevHypo(nonTermIndexMap[phrasePos]);
      const DALMChartState* prevState =
        static_cast<const DALMChartState*>(prevHypo->GetFFState(featureID));

      size_t prevTargetPhraseLength = prevHypo->GetCurrTargetPhrase().GetSize();

      EvaluateNonTerminal(
        word, hypoScore,
        newState, state,
        prefixFragments, prefixLength,
        prevState, prevTargetPhraseLength
      );
    }
  }
  hypoScore = TransformLMScore(hypoScore);
  hypoScore -= hypo.GetTranslationOption().GetScores().GetScoresForProducer(this)[0];

  // assign combined score to score breakdown
  if (OOVFeatureEnabled()) {
    std::vector<float> scores(2);
    scores[0] = hypoScore;
    scores[1] = 0.0;
    out->PlusEquals(this, scores);
  } else {
    out->PlusEquals(this, hypoScore);
  }

  return newState;
}

bool LanguageModelDALM::IsUseable(const FactorMask &mask) const
{
  return mask[m_factorType];
}

void LanguageModelDALM::CreateVocabMapping(const std::string &wordstxt)
{
  InputFileStream vocabStrm(wordstxt);

  std::vector< std::pair<std::size_t, DALM::VocabId> > vlist;
  string line;
  std::size_t max_fid = 0;
  while(getline(vocabStrm, line)) {
    const Factor *factor = FactorCollection::Instance().AddFactor(line);
    std::size_t fid = factor->GetId();
    DALM::VocabId wid = m_vocab->lookup(line.c_str());

    vlist.push_back(std::pair<std::size_t, DALM::VocabId>(fid, wid));
    if(max_fid < fid) max_fid = fid;
  }

  for(std::size_t i = 0; i < m_vocabMap.size(); i++) {
    m_vocabMap[i] = m_vocab->unk();
  }

  m_vocabMap.resize(max_fid+1, m_vocab->unk());
  std::vector< std::pair<std::size_t, DALM::VocabId> >::iterator it = vlist.begin();
  while(it != vlist.end()) {
    std::pair<std::size_t, DALM::VocabId> &entry = *it;
    m_vocabMap[entry.first] = entry.second;

    ++it;
  }
}

DALM::VocabId LanguageModelDALM::GetVocabId(const Factor *factor) const
{
  std::size_t fid = factor->GetId();
  return (m_vocabMap.size() > fid)? m_vocabMap[fid] : m_vocab->unk();
}

void LanguageModelDALM::SetParameter(const std::string& key, const std::string& value)
{
  if (key == "factor") {
    m_factorType = Scan<FactorType>(value);
  } else if (key == "order") {
    m_nGramOrder = Scan<size_t>(value);
  } else if (key == "path") {
    m_filePath = value;
  } else {
    LanguageModel::SetParameter(key, value);
  }
  m_ContextSize = m_nGramOrder-1;
}

void LanguageModelDALM::EvaluateTerminal(
  const Word &word,
  float &hypoScore,
  DALMChartState *newState,
  DALM::State &state,
  DALM::Fragment *prefixFragments,
  unsigned char &prefixLength) const
{

  DALM::VocabId wid = GetVocabId(word.GetFactor(m_factorType));
  if (newState->LargeEnough()) {
    float score = m_lm->query(wid, state);
    hypoScore += score;
  } else {
    unsigned char prevLen = state.get_count();
    float score = m_lm->query(wid, state, prefixFragments[prefixLength]);

    if(score > 0) {
      hypoScore -= score;
      newState->SetAsLarge();
    } else if(state.get_count()<=prefixLength) {
      hypoScore += score;
      prefixLength++;
      newState->SetAsLarge();
    } else {
      hypoScore += score;
      prefixLength++;
      if(state.get_count() < std::min(prevLen+1, (int)m_ContextSize)) {
        newState->SetAsLarge();
      }
      if(prefixLength >= m_ContextSize) newState->SetAsLarge();
    }
  }
}

void LanguageModelDALM::EvaluateNonTerminal(
  const Word &word,
  float &hypoScore,
  DALMChartState *newState,
  DALM::State &state,
  DALM::Fragment *prefixFragments,
  unsigned char &prefixLength,
  const DALMChartState *prevState,
  size_t prevTargetPhraseLength
) const
{

  const unsigned char prevPrefixLength = prevState->GetPrefixLength();
  const DALM::Fragment *prevPrefixFragments = prevState->GetPrefixFragments();

  if(prevPrefixLength == 0) {
    newState->SetAsLarge();
    hypoScore += m_lm->sum_bows(state, 0, state.get_count());
    state = prevState->GetRightContext();
    return;
  }
  if(!state.has_context()) {
    newState->SetAsLarge();
    state = prevState->GetRightContext();
    return;
  }
  DALM::Gap gap(state);
  unsigned char prevLen = state.get_count();

  // score its prefix
  for(size_t prefixPos = 0; prefixPos < prevPrefixLength; prefixPos++) {
    const DALM::Fragment &f = prevPrefixFragments[prefixPos];

    if (newState->LargeEnough()) {
      float score = m_lm->query(f, state, gap);
      hypoScore += score;

      if(!gap.is_extended()) {
        state = prevState->GetRightContext();
        return;
      } else if(state.get_count() <= prefixPos+1) {
        state = prevState->GetRightContext();
        return;
      }
    } else {
      DALM::Fragment &fnew = prefixFragments[prefixLength];
      float score = m_lm->query(f, state, gap, fnew);
      hypoScore += score;

      if(!gap.is_extended()) {
        newState->SetAsLarge();
        state = prevState->GetRightContext();
        return;
      } else if(state.get_count() <= prefixPos+1) {
        if(state.get_count() == prefixPos+1 && !gap.is_finalized()) {
          prefixLength++;
        }
        newState->SetAsLarge();
        state = prevState->GetRightContext();
        return;
      } else if(gap.is_finalized()) {
        newState->SetAsLarge();
      } else {
        prefixLength++;
        if(state.get_count() < std::min(prevLen+1, (int)m_ContextSize)) {
          newState->SetAsLarge();
        }

        if(prefixLength >= m_ContextSize) newState->SetAsLarge();
      }
    }
    gap.succ();
    prevLen = state.get_count();
  }

  // check if we are dealing with a large sub-phrase
  if (prevState->LargeEnough()) {
    newState->SetAsLarge();
    //if(prevPrefixLength < prevState->GetHypoSize()) {
    hypoScore += m_lm->sum_bows(state, prevPrefixLength, state.get_count());
    //}
    // copy language model state
    state = prevState->GetRightContext();
  } else {
    m_lm->set_state(state, prevState->GetRightContext(), prevPrefixFragments, gap);
  }
}

}